Auto Clustering Prediction Models for Retail Demand Forecasting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sales forecasting systems in the retail industry face challenges in producing accurate week-by-week forecasts due to the complexity of factors influencing demand, such as promotions, price, and seasonality, often resulting in inadequate historical data quality leading to over/understocking and incorrect planning.
Innovation Solution
The implementation of an auto-clustering prediction model that aggregates sales data across products and locations, extracts relevant features, and trains cluster models to estimate promotion effects, allowing for time-sensitive demand forecasting by mapping features to trained cluster models to predict future demand.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional causal demand models are used to forecast sales, then the model structure is simple and easy to implement, but the forecast accuracy deteriorates due to inadequate historical data quality and inability to separate effects of multiple demand variables
Solution Approach 1:
The patent segments the historical sales data into multiple clusters based on similarity in demand patterns and features. Each cluster represents a distinct group of products with similar characteristics, allowing the model to capture nuanced relationships between demand variables and sales outcomes. This segmentation enables more accurate forecasting by treating heterogeneous products differently rather than applying a uniform model.
Solution Approach 2:
The patent transforms the traditional causal demand model by changing the parameters from fixed regression coefficients to cluster-specific trained models. Each cluster has its own set of demand parameters that are optimized based on historical data within that cluster, allowing the system to adapt to different product characteristics and demand patterns while maintaining the overall model structure.
2Loss of information
If regression is performed on demand models to separate effects of multiple demand variables, then the relationship between variables can be determined, but the quality of forecasts deteriorates when historical data is inadequate
Solution Approach 1:
The patent merges data from multiple products into clusters based on similarity in demand patterns. By combining historical data across products within the same cluster, the system effectively increases the amount of training data available for each cluster model, improving the reliability of parameter estimation even when individual product histories are short or incomplete.
Solution Approach 2:
The patent creates cluster-specific copies of the demand model, where each cluster has its own trained parameters and relationships. This allows the system to leverage patterns learned from aggregated cluster data and apply them to individual products within the cluster, effectively copying successful forecasting patterns from similar products to improve forecasts for products with limited historical data.
3Measurement precision
If cluster models are trained for each product/location combination, then the forecast accuracy improves by accounting for unique impacts, but the computational complexity and data requirements increase significantly
Solution Approach 1:
The patent segments products into clusters based on similarity in demand patterns and features, rather than creating separate models for each individual product. This segmentation reduces the number of models that need to be trained and maintained, decreasing data requirements while still capturing product-specific characteristics through the cluster assignments.
Solution Approach 2:
The patent creates universal cluster models that serve multiple products within each cluster. Each cluster model functions as a universal predictor for all products in that cluster, allowing the system to achieve product-specific forecasting accuracy without requiring separate data sets for each product. The cluster models are multi-functional, handling forecasting for multiple products simultaneously.
Data Source
AI summary
Embodiments predict future demand for a first product by receiving historical sales data for an aggregate products/locations level, the historical sales data including a plurality of sales data points, including sales data points for the first product at each of a plurality of locations. Embodiments extract a plurality of different types of features related to sales of each of the products and generate a plurality of clusters of sales data points based on the plurality of different types of features. Embodiments train each of the clusters to generate a plurality of trained cluster models including promotion effects per cluster. For a particular time period, a particular location and the first product, embodiment identify the features for the time period and map to one of the trained cluster models to fetch the promotion effects for the time period. Embodiments then use the promotion effects to forecast demand for the first product.